I build backend services in Python, and lately most of them have a model somewhere inside: semantic search over private documents, tools that let an assistant query scientific databases, and the training code around small decision models. I like work I can measure, so my projects usually come with a benchmark and an honest comparison against what already exists.

An MCP server that gives Claude Desktop scientific research tools: protein lookup in UniProt, PubMed search and reading through PubTator3, and semantic search over a local PDF library in Portuguese and English, with OCR for scanned files. The documents and the vector index never leave the machine.

Python MCP RAG Qdrant EasyOCR

Rigid-body alignment and multi-copy packing of protein structures into SAXS envelopes, using ICP with a bidirectional volumetric penalty and a score in Ångströms. Benchmarked on 50 SASBDB entries against ATSAS cifsup: it improves on unconstrained ICP in every metric and is a comparable open-source alternative to the NSD mode.

Python NumPy SciPy Biopython

A modular RAG backend for chatting with private PDF and TXT files through a local LLM: JWT authentication, asynchronous document processing with status tracking, sliding-window chunking, and cosine search in pgvector.

FastAPI PostgreSQL + pgvector Ollama Docker